Just records the input whenever the test fails, and have an option to re-run a test, with those inputs. The tests are still flaky but it doesn't matter if we can reproduce the results.

You don't even have to record all inputs. You could just use a random seed, and record the seed at each run. Test become deterministic again simply by fixing the seed (typically to a value that caused the previous run to fail).

If you want to have the build (test) result feedback to save the input, that's an ok goal but I think it's hard to pull off: now you have your build making commits to source control (causing another build), or you have test data stored somewhere else not in source (when really, it should be part of source).

Here's a link that talks about the above in more detail - https://software.rajivprab.com/2019/04/28/rethinking-softwar...

I use a seeded random number generator. During development I seed it randomly. When crashes happen I output both the seed and iteration so I can hardcode that seed and reproduce the problem. And then in the CI I’ll often fix the seed so the test suite is stable.

But it’s also good practice to run fuzzers outside the context of a CI. You want your CI tests to be fast and stable. Fuzzer tests can also be slow and thorough. Foundationdb has dedicated hardware that runs their fuzzer 24/7 and spits out any errors it finds. If that fuzzer finds problems it doesn’t break the build. But someone will take a look, reproduce the problem, hopefully convert the bug into a unit test and fix the issue.

On a separate note, if the fuzzy tests find something wrong, the data should definitely be integrated in the stable test suite to ensure non regression.

I guess another controversial opinion: this is a problem with the idea of CI (or at least how we work with it and what we expect to get out of it), rather than the idea of randomized tests.

Flaky tests are an indicator of poor code: maybe it's your actual code, maybe it's a bug in the test code, or maybe it's external dependency + lack of error handling in the test code but there's a problem somewhere.

I did admit this was controversial! But it fits in with a more general view that there are a lot of tools which make good servants but poor masters. Auto-builders are a good thing, partly because (as you say) they can help to clarify what is required to make a build, and partly because (especially for dependency-heavy software, which seems to be the norm nowadays) they can help catch things quickly when the dependencies shift beneath you. Making them a hard gate on releases seems a little too close to making the tooling your master, though.

(Somewhat separately, I also worry about CI acts as a hiding place for complexity. Sometimes it reaches that point where nobody knows how to make a build without the CI tool any more. Then local testing and debugging becomes difficult.)

I do also agree that CI can hide complexity, but that's true of any tool as well as non-CI builds. Compared to a human running things, it's at least somewhat self-documenting due to being scripted.